Meta has released the AI agent Muse Code for programming on macOS and Linux.

Meta announced the launch of Muse Code, a terminal-based AI agent for software development. The tool is available in beta for macOS and Linux, and installation is performed via the command line. At the core of Muse Code is Meta's new model, Muse Spark 1.2.

Muse Code is designed to handle complex tasks in large repositories, including planning changes, writing code, and verifying results. The agent can coordinate multiple parallel background agents within a single task, which, according to developers, accelerates the resolution of complex problems and reduces the need for user involvement.

Background agents remain active throughout the entire session rather than being recreated for each individual task, allowing the system to avoid redundant information gathering. To track actions, Muse Code maintains a local event log that records all model calls, tool executions, confirmations, and applied edits. This enables the agent to resume work after a failure and continue the task from where it was interrupted.

The tool comes with a set of built-in commands: "plan" converts a task into a work plan with confirmation at each stage, "grill" stress-tests this plan, and "goal" directs the agent toward achieving the set objective.

Muse Spark 1.2 is an update to Muse Spark 1.1, focused on programming. It features improvements in code generation, debugging complex errors, understanding project structure, and handling long development cycles. Muse Spark 1.2 was trained jointly with Muse Code, ensuring compatibility and better performance when used together.

Muse Spark 1.2 is already available within Muse Code and via the Meta Model API, with API access expanded to additional regions.

Source: research.meta.ai